Improved Lane Detection Network Based on Instance Segmentation

Abstract

In recent years, with the development of artificial intelligence technology, the traditional automobile industry is closely combined with computer vision technology. Intelligent transportation is becoming an important research topic. Lane detection, as one of basic components in computer vision, is the key technology to achieve this goal. Generally, there are two main branches for lane detection. The traditional approach often needs edge detection, lane fitting and tracking. These methods greatly depend on hand-crafted features which conduct poor capacity of resisting disturbance and stabilization. The deep learning approach can extract lane features automatically which improves accuracy and robustness in various environment conditions. Based on deep learning methods, we proposed a new lane detection method, which uses semantic segmentation powerful multi-level and hybrid features encoder and utilizes binary segmentation and pixel embedding decoders. By using multilevel features, the proposed network can predict reliable lane recognition effectively and precisely. The experimental results with challenging public datasets show distinct performance improvement over other methods and demonstrate the effect of our method for lane detection.